Recent advancements in Large Language Models (LLMs) have led to their adaptation in various domains as conversational agents. We wonder: can personality tests be applied to these agents to analyze their behavior, similar to humans? We introduce TRAIT, a new benchmark consisting of 8K multi-choice questions designed to assess the personality of LLMs. TRAIT is built on two psychometrically validated small human questionnaires, Big Five Inventory (BFI) and Short Dark Triad (SD-3), enhanced with the ATOMIC-10X knowledge graph to a variety of real-world scenarios. TRAIT also outperforms existing personality tests for LLMs in terms of reliability and validity, achieving the highest scores across four key metrics: Content Validity, Internal Validity, Refusal Rate, and Reliability. Using TRAIT, we reveal two notable insights into personalities of LLMs: 1) LLMs exhibit distinct and consistent personality, which is highly influenced by their training data (e.g., data used for alignment tuning), and 2) current prompting techniques have limited effectiveness in eliciting certain traits, such as high psychopathy or low conscientiousness, suggesting the need for further research in this direction.

Unsupervised learning objectives like autoregressive and masked language modeling constitute a significant part in producing pre-trained representations that perform various downstream applications from natural language understanding to conversational tasks. However, despite impressive generative capabilities of recent large language models, their abilities to capture syntactic or semantic structure within text lag behind. We hypothesize that the mismatch between linguistic performance and competence in machines is attributable to insufficient learning of linguistic structure knowledge via currently popular pre-training objectives. Working with English, we show that punctuation restoration as a learning objective improves performance on structure-related tasks like named entity recognition, open information extraction, chunking, and part-of-speech tagging. Punctuation restoration results in ▲≥2%p improvement in 16 out of 18 experiments, across 6 out of 7 tasks. Our results show that punctuation restoration is an effective learning objective that can improve structure understanding and yield a more robust structure-aware representations of natural language in base-sized models.

The advent of scalable deep models and large datasets has improved the performance of Neural Machine Translation (NMT). Knowledge Distillation (KD) enhances efficiency by transferring knowledge from a teacher model to a more compact student model. However, KD approaches to Transformer architecture often rely on heuristics, particularly when deciding which teacher layers to distill from. In this paper, we introduce the “Align-to-Distill” (A2D) strategy, designed to address the feature mapping problem by adaptively aligning student attention heads with their teacher counterparts during training. The Attention Alignment Module (AAM) in A2D performs a dense head-by-head comparison between student and teacher attention heads across layers, turning the combinatorial mapping heuristics into a learning problem. Our experiments show the efficacy of A2D, demonstrating gains of up to +3.61 and +0.63 BLEU points for WMT-2022 De→Dsb and WMT-2014 En→De, respectively, compared to Transformer baselines.The code and data are available at https://github.com/ncsoft/Align-to-Distill.

Generating diverse and consistent responses is the ultimate goal of a persona-based dialogue. Although many studies have been conducted, the generated responses tend to be generic and bland due to the personas’ limited descriptiveness. Therefore, it is necessary to expand the given personas for more attractive responses. However, indiscriminate expansion of personas threaten the consistency of responses and therefore reduce the interlocutor’s interest in conversation. To alleviate this issue, we propose a consistent persona expansion framework that improves not only the diversity but also the consistency of persona-based responses. To do so, we define consistency criteria to avoid possible contradictions among personas as follows: 1) Intra-Consistency and 2) Inter-Consistency. Then, we construct a silver profile dataset to deliver the ability to conform with the consistency criteria to the expansion model. Finally, we propose a persona expansion model with an encoder-decoder structure, which considers the relatedness and consistency among personas. Our experiments on the Persona-Chat dataset demonstrate the superiority of the proposed framework.

A persona-grounded dialogue model aims to improve the quality of responses to promote user engagement. However, because the given personas are mostly short and limited to only a few informative words, it is challenging to utilize them to generate diverse responses. To tackle this problem, we propose a novel persona expansion framework, Concept-based Persona eXpansion (CPX). CPX takes the original persona as input and generates expanded personas that contain conceptually rich content. We constitute CPX with two task modules: 1) Concept Extractor and 2) Sentence Generator. To train these modules, we exploit the duality of two tasks with a commonsense dataset consisting of a concept set and the corresponding sentences which contain the given concepts. Extensive experiments on persona expansion and response generation show that our work sufficiently contributes to improving the quality of responses in diversity and richness.

A lack of consistency in terminology translation undermines quality of translation from even the best performing neural machine translation (NMT) models, especially in narrow domains like literature, medicine, and video game jargon. Dictionaries containing terminologies and their translations are often used to improve consistency but are difficult to construct and incorporate. We accompany our submissions to the WMT ‘23 Terminology Shared Task with a description of our experimental setup and procedure where we propose a framework of terminology-aware machine translation. Our framework comprises of an automatic terminology extraction process that constructs terminology-aware machine translation data in low-supervision settings and two model architectures with terminology constraints. Our models outperform baseline models by 21.51%p and 19.36%p in terminology recall respectively on the Chinese to English WMT’23 Terminology Shared Task test data.

One of the challenges of developing a summarization model arises from the difficulty in measuring the factual inconsistency of the generated text. In this study, we reinterpret the decoder overconfidence-regularizing objective suggested in (Miao et al., 2021) as a hallucination risk measurement to better estimate the quality of generated summaries. We propose a reference-free metric, HaRiM+, which only requires an off-the-shelf summarization model to compute the hallucination risk based on token likelihoods. Deploying it requires no additional training of models or ad-hoc modules, which usually need alignment to human judgments. For summary-quality estimation, HaRiM+ records state-of-the-art correlation to human judgment on three summary-quality annotation sets: FRANK, QAGS, and SummEval. We hope that our work, which merits the use of summarization models, facilitates the progress of both automated evaluation and generation of summary.